resume-agent-sessions

name	resume-agent-sessions
description	Use when the user wants to RESUME or CONTINUE prior coding-agent work — "resume my codex session", "continue where I left off in <project>", "what sessions ran in this folder", "pick up that grok conversation", or any ask for session/resume ids across agents (Claude Code, Codex, Grok, Cursor Agent, …). Surfaces every agent's recent sessions under a path with the exact resume command for each.

Resume any coding agent's session from a path

Why this skill exists

Every coding agent keeps its own session history with its own resume UX — claude --resume, codex resume, grok --continue — and none of them can see the others. StackUnderflow indexes all of them, so one local CLI call answers "what was running here, in any tool, and how do I get back into it".

When this skill fires

• "Resume my codex session in ~/dev/foo" / "continue that grok run"
• "What sessions have run in this folder?" (across agents)
• "Give me the resume id for the claude session from Tuesday"
• The user switches projects and wants to pick up prior work in ANY tool.

Do NOT fire when the user wants this conversation to recall past context — that's the stackunderflow memory namespace (decisions / file / ask), not session resumption.

Step 1: one call

stackunderflow resume <path> --json

Add -p/--provider <agent> (repeatable; unambiguous prefix ok) when the user names a specific tool — "resume my codex session" → stackunderflow resume <path> -p codex --json.

<path> defaults to cwd. Matching is bidirectional: standing inside a project finds it, and a workspace folder (e.g. ~/dev/my_workspace/) lists every project underneath it. Underscore/hyphen path quirks are handled (matching is slug-space).

Envelope (stackunderflow.resume/1):

{
  "schema": "stackunderflow.resume/1",
  "path": "/abs/path",
  "providers": [
    {"provider": "codex",
     "resume": {"command": "codex resume {session_id}", "scope": "session"},
     "sessions": [
       {"session_id": "…", "last_ts": "…", "message_count": 151,
        "project": "-slug", "project_path": null,
        "resume_command": "codex resume 019f…"}
     ]}
  ]
}

Step 2: act on the shape

• resume.scope == "session" → each session carries a ready resume_command. Present it to the user (with when/size context so they pick the right one).
• resume.scope == "latest" (e.g. grok) → there is no per-id resume; tell the user to run the provider-level command from that project's directory.
• resume == null → the tool has no verified resume command for that agent; give the session id and never invent a flag.

Rules

• Present commands; don't execute them. Resume commands launch interactive TUIs — running one inside an agent session hangs it. Hand the command to the user (or write it to their terminal input if the harness supports that).
• All data is local SQLite, read-only, <100 ms. Run liberally.
• Human-readable form (no --json) is fine to paste verbatim when the user just asks "what can I resume here?".